Branch Predicting with Sparse Distributed Memories
share
Modern processors rely heavily on speculation to keep the pipeline filled and consequently execute and commit instructions as close to maximum capacity as possible. To improve instruction-level parallelism, the processor core needs to fetch and decode multiple instructions per cycle and has come to rely on incredibly accurate branch prediction. However, this comes at cost of the increased area and complexity which is needed for modern high accuracy branch predictors. The key idea described in this work is to use hyperdimensional computing and sparse distributed memory principles to create a novel branch predictor that can deliver complex predictions for a fraction of the current area. Sparse distributed memories can store vast amounts of data in a compressed manner, theoretically enabling branch histories larger and more precise than the branch predictors used today to be stored with equal or smaller area footprint. Furthermore, as all the data is in a hashed format and due to the nature of the hashing scheme used, it is inherently harder to manipulate with known side-channel attacks. We describe our proof-of-concept and evaluate it against a state-of-the-art academic TAGE predictor. Our experiments are conducted on realistic synthetic branch predictor patterns and the Championship Branch Prediction traces and show competitive accuracy. Finally, we describe techniques that can be used to solve some of the challenges of processing with hyperdimensional vectors in order to deliver timely predictions.
READ FULL TEXT